> I would like you about your experience about running amber extensive
> simulations on a cloud (amazon or any other) vs purchasing a new GPU x 4
> machine. I have no experience at all with clouds.
Cloud will be considerably more expensive, especially if you use a GPU in
the cloud. CPU is 2-3x the cost, and GPU much higher (since they mostly do
not deploy commodity cards). However, if you workload doesn't require 24/7 use
of the GPUs or is very bursty you could consider the cloud.
> How difficult it is to deal with implementation and compilation issues
> on a cloud?
It is not as easy as the Cloud providers make it out to be, but if you
containerized AMBER it could be easily run. There are also vendors like
rescale that will automanage the software, cloud credits, etc for you for
a slight premium.
> My comparison should be based on purchasing a new GPU X 4 machine, for
> which I would invest let’s say about 6000$, and comparing it to running
> on a cloud instead.
Let's see, just based on published rates for GPU instances on AWS (noting
you can probably get slightly lower prices) for current generation GPUs,
cost is 0.188 per hour, so $6K gets you ~32K hours or ~3.6 years on a
single GPU. Your 4-node GPU will give you 4 GPUs for ~5 years each and
this is not including data movement charges (although the $6K on-prem
doesn't factor in power costs, etc).
[Addendum, now I cannot find the 0.188 per hour price and instead see
$0.75 which means 8,000 hours or just under 1-year on a single GPU.]
--tec3
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Sep 18 2019 - 14:30:01 PDT
